MMOOZZGG.com

AI investment

Build or buy AI: how to choose and not overpay

5 min read3 sources

Build or buy AI? Large companies in Russia more often buy: 78% of them use ready AI solutions from outside partners, per a 2025 industry study. The better option is the one that costs less over the life of the solution and does not tie the company to one supplier.

On this page · 5 sections

What large companies do

A 2025 industry study surveyed 150 technical directors of large Russian companies from 16 industries. 78% use ready AI solutions from outside partners, and 71% fine-tune ready models instead of a full training cycle.

So for most companies, the choice is not between your own model and someone else’s. It is between a ready service and your own solution built on a ready model.

A mid-size company usually has no AI development team of its own. For it, the choice looks like this: a subscription service, or a solution that a contractor builds on a ready model.

Worldwide, there is a visible move the other way. In a 2026 McKinsey survey, 32% of organizations dropped the purchase of at least one product because they can build it themselves with AI agents.

That is why the build-or-buy question is decided for each process, not once for the whole company.

When to build

Your own solution pays back when the process sets you apart from competitors, no ready product exists, or the data cannot go to an outside service.

1. The process is unique: your own pricing logic, production, or purchasing with special rules.

2. The data is the value of the company, and you cannot hand it to an outside service.

3. Ready services cover less than half of the task, and adapting them costs more than your own solution.

4. The solution must sit deep inside your accounting, CRM and production systems, and a ready service works only through its own windows.

Before you decide to build, count everything: development, support after launch, model fees and the person who answers for the solution inside the company.

Building does not mean doing it all yourself. Per MIT NANDA 2025, projects with an outside partner reached working use in about 67% of cases, against 33% for internal builds. These are survey data, not an exact measurement.

A company rarely needs to build its own neural network from scratch: 71% of large companies in the Russian study take a ready model and fine-tune it on their own data.

When to buy

A ready solution wins when the task is typical: answers to common questions, client booking, call transcription, document checks. Such services already exist, their price is known, and launch is faster than development.

That is what 78% of large Russian companies in the 2025 industry study do. They buy a ready product if it covers the process and costs less than their own development over two or three years.

Check three things before a purchase: where the data is processed, whether you can export it if you leave, and what happens if the supplier raises the price or switches the service off.

A ready solution also needs work: connect it to the CRM and accounting, write the rules, name an owner. Add these costs to the comparison, or the purchase looks cheaper than it is.

The opposite mistake: a contractor offers development where a ready service is enough. How to spot a contractor who sells development is covered separately.

Your own AI on your own servers

Between a ready service and your own development there is a third path: a ready open model on the company’s servers. Data does not leave, and nobody outside can switch the service off.

It is the same approach as 71% of large companies in the 2025 industry study: a ready model fine-tuned for your own tasks. The difference is where it runs.

You need this option if you process personal data or trade secrets and must keep data in Russia under 152-FZ, Russia’s personal data law.

Companies also choose their own servers because of the risk of a shutdown: access to a foreign service can be closed by another country’s decision, without warning.

The downside of this path: servers and model updates fall on the company or its contractor. For a small business this may cost more than a ready service that processes data in Russia.

The cost is made up of servers, setup and support. How sovereign AI works and who needs it is covered in a separate section.

How to decide in a week

You can make a decision on one process in a week if you go step by step.

1. Describe the process and count what it costs now: staff hours, salary with taxes, errors.

2. Find ready solutions on the market and their price per year.

3. Estimate your own development: timeline, team, support after launch.

4. Check the data: which of it is personal and where you may process it under 152-FZ.

5. Compare the options by price over two or three years and by dependence risk, then decide.

The result of the week: a written decision for the process (build, buy or leave it alone), its price over two or three years and a pilot metric.

Count the share of failures: per MIT NANDA, internal builds reached working use in 33% of cases. That is why a pilot with a metric comes before full rollout.

The AI solutions map takes these steps for all the company’s processes at once and shows where to build, where to buy and where AI is not needed.

Sources and fact-checking

Checked by the editors on 14 October 2026.

  1. Yandex and Yakov and Partners: a study of AI in Russia, 8 December 2025 (yandex.ru)
  2. McKinsey, The state of AI in 2026, August 2026 (mckinsey.com)
  3. MIT Media Lab, Project NANDA: The GenAI Divide, State of AI in Business 2025, July 2025 (mlq.ai)

Found a mistake? Tell us

Publications

Contact us